A method of optically implementing a restricted boltzmann machine

Abstract

The application discloses a method for realizing a restricted Boltzmann machine optically. The application can realize analog computing of a restricted Boltzmann machine and optical Gibbs sampling, and has the advantages of wide application range, fast information transmission, simple structure, low cost and fast computing. The application uses a coherent wide-spectrum light source as a signal input, uses a grating or other light splitting device to split the light field, and then the light field is incident on a modulator. Then, the spin, interaction and magnetic field parameters are encoded on the light wavefronts at different positions through a time or space light modulator. An optical Fourier transform is performed by using a time or space lens system. The light intensity after the Fourier transform is measured by using a detector. Finally, the optical Gibbs sampling is realized by calculating the difference between the light intensities measured twice, the computing complexity is reduced, and the computing efficiency is improved. The method provided by the application has important application prospects in the field of optical neural networks and the like, and can realize applications such as generated content and classification identification, and is convenient for integration into an optical chip and the like.

Description

Technical Field

[0001] This invention relates to the fields of optical computing and machine learning, and in particular to a method for optically implementing a restricted Boltzmann machine and Gibbs sampling. Background Technology

[0002] In recent years, generative artificial intelligence (AI) has developed rapidly, but processing complex information requires a large amount of computing resources and time. Currently, the miniaturization of electronic devices is approaching its physical limits, and electronic computing faces a dual bottleneck of computing power and energy consumption. Traditional electronic computing devices, such as GPUs, are gradually failing to meet the ever-increasing computing power demands of generative AI. Especially recently, generative AI, represented by the GPT model, relies on large language models with hundreds of millions of parameters, consuming a significant amount of energy, storage space, and computing resources during training. Therefore, the need for new computing architectures in generative AI is becoming increasingly urgent. Optical computing can solve these problems, achieving ultra-high-speed, low-energy computation using photons. Different wavelengths of light can carry different information, enabling wavelength division multiplexing; stable transmission is also possible under high-frequency modulation; furthermore, optical transmission offers excellent confidentiality and is unaffected by magnetic field interference; Fourier transforms are performed at the speed of light, greatly increasing the speed of two-dimensional matrix operations. Therefore, optical computing provides a new approach to complex and rapid computation.

[0003] Restricted Boltzmann machines (RBMs) are a type of stochastic generative neural network widely used in tasks such as deep learning, image classification, and image generation. However, in traditional electronic computing architectures, training an RBM model requires multiple iterations of Markov chains for Gibbs sampling, which is computationally expensive, especially on large datasets. Therefore, implementing Gibbs sampling using optical methods can accelerate computation and save resources. However, current systems based on optical modulators cannot simulate the computation of RBMs and Gibbs sampling. The method proposed in this invention can realize RBMs and Gibbs sampling, increasing the capacity for information transmission and improving the speed of matrix operations, providing a new approach for the research of optical generative neural networks. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for optically realizing a restricted Boltzmann machine, which can also achieve optical Gibbs sampling.

[0005] The objective of this invention is achieved through the following technical solution: a method for optically realizing a restricted Boltzmann machine, based on a non-von Neumann architecture, to achieve optical Gibbs sampling; using a coherent broadband light source as the incident light, and using a beam splitter such as a grating to split the light field by wavelength before incident on an optical modulator; encoding spin, interaction, and magnetic field parameters into optical signals at different locations using a time- or space-based optical modulator; performing an optical Fourier transform through a lens system, and measuring the light intensity after the Fourier transform using a detector; finally, calculating the difference between the two measured light intensities to achieve optical Gibbs sampling, thereby reducing computational complexity and improving computational efficiency. This invention reduces the computational complexity of Gibbs sampling from O(N) to O(1), significantly improving computational speed, while also possessing advantages such as wide applicability, fast information transmission, simple structure, and low cost. This invention has significant application prospects in fields such as optical neural networks, enabling applications such as content generation and classification recognition. The experimental optical path is simple, requiring only one modulator, further reducing costs; the device size can be designed on the wavelength scale, facilitating integration into specific applications such as optical chips.

[0006] Furthermore, it is possible to achieve a Hamiltonian of... A restricted Boltzmann machine, where v = {v1,...,v...} i ,...,v K} represents the visible layer spin, h = {h1,...,h} j ,...,h K} represents the hidden layer spin, W ij visible layer spin v i With hidden layer spin h j The interaction between them, a i visible layer spin v i The magnetic field parameters, b i For hidden layer spin h i The magnetic field parameters are given by K, where K is the total number of wavelengths.

[0007] Furthermore, based on a non-von Neumann architecture, optical Gibbs sampling is achieved, encoding spin, interaction, and magnetic field parameters onto the optical modulator, eliminating the need for additional parameter storage and enabling in-memory computing.

[0008] Furthermore, in optical Gibbs sampling, the spatial light modulator needs to be divided into three regions to encode the hidden layer spin or visible layer spin, interaction, and magnetic field parameters, respectively.

[0009] Furthermore, based on the spin, interaction, and magnetic field parameters of the visible and hidden layers, optical signals at different locations are modulated by an optical modulator, which includes, but is not limited to, an electro-optic modulator and a spatial optical modulator, and the modulation method includes, but is not limited to, amplitude modulation or phase modulation; when all magnetic field parameters are zero, the modulator can modulate the optical signal only based on spin and interaction.

[0010] Furthermore, an optical Fourier transform is performed using a lens system, which includes, but is not limited to, ordinary lenses, micro / nano structure lenses, and time lenses.

[0011] Furthermore, after passing through the lens system, a detector is used to measure the intensity of the light after Fourier transform. The detector includes, but is not limited to, a photodetector and an optical camera.

[0012] Furthermore, by calculating the difference in light intensity between the two detections, the Gibbs sampling probability can be calculated, reducing the computational complexity of Gibbs sampling from O(N) to O(1), which greatly improves the computational speed.

[0013] Furthermore, based on the optical Gibbs sampling method, the training time required for the restricted Boltzmann machine can be reduced, thereby improving computational efficiency.

[0014] The beneficial effects of this invention are as follows: This invention proposes a method for implementing a restricted Boltzmann machine and Gibbs sampling, reducing computational complexity and improving computational efficiency. Simultaneously, the system can be designed as a subwavelength structure, facilitating integration onto optical chips. In practical applications, the optical restricted Boltzmann machine proposed in this invention can be applied to artificial intelligence tasks such as content generation and classification, possessing advantages such as wide applicability, fast information transmission, simple structure, and low cost. Attached Figure Description

[0015] Figure 1 An optical path diagram for an embodiment implemented using a phase-type spatial light modulator;

[0016] Figure 2 Example diagram of training and generating content for a restricted Boltzmann machine;

[0017] Figure 3 To implement the coding method on the optical Gibbs sampling spatial light modulator;

[0018] Figure 4 The process of generating images for an optically confined Boltzmann machine;

[0019] Figure 5 This is an optical path diagram of an embodiment implemented using an electro-optic modulator. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] like Figure 1 The diagram shown illustrates the optical path of an embodiment. It includes a supercontinuum laser, a grating, a cylindrical lens, a phase-type spatial light modulator, a lens, and a detector. The horizontal direction of the phase-type spatial light modulator is defined as the x-axis, and the vertical direction as the y-axis. The phase-type spatial light modulator is located at the rear focal plane of the cylindrical lens and the front focal plane of the lens, while the detector is located at the rear focal plane of the lens. The total number of wavelengths is set to K. The supercontinuum laser emits collimated light, which, after passing through a beam expander, is incident on the grating. Due to the beam splitting characteristics of the grating, different wavelengths (i.e., K wavelengths) exit at different angles along the x-direction. The beam then undergoes a one-dimensional optical Fourier transform along the x-direction through the cylindrical lens, collimating and imaging the beam onto the phase-type spatial light modulator. The resulting beam is partially coherent; the vertical direction, with the same wavelength, allows interference, while the horizontal direction, with its varying wavelengths, prevents interference. Spin, interaction, and magnetic field parameters are then encoded into the phase-type spatial light modulator, thereby altering the intensity and phase distribution of the wavefront. Coherent light beams of the same wavelength undergo an optical Fourier transform through a lens, while light beams of different wavelengths combine their light field intensities at the lens's focal plane. A detector then measures the light field intensity at the lens's focal plane. Finally, a computer provides feedback on the detected intensity and further modulates the light field.

[0022] like Figure 2 The image shown is an example diagram illustrating the training and content generation of a Restricted Boltzmann Machine. Figure 3 Taking K=3 as an example. Where v is the visible layer spin, v = {v1,...,v} i ,...,v K}, where h is the hidden layer spin, h = {h1,...,h} j ,...,h K}, W ij visible layer spin v i With hidden layer spin h j The interaction between them, a i visible layer spin v i The magnetic field parameters, b i For hidden layer spin h i The magnetic field parameters. The Hamiltonian of a confined Boltzmann machine is expressed as... Interaction and magnetic field parameters are obtained by training a restricted Boltzmann machine using images of the same type and maximizing the likelihood function. Different generated images of the same type can then be obtained through Gibbs sampling iterations on the trained restricted Boltzmann machine.

[0023] like Figure 3As shown, this is an encoding method for optical Gibbs sampling. The phase-type spatial light modulator is divided into three regions along the y-axis, each encoding different parameters. During the process of sampling the hidden layer spin from the visible layer spin, the top region I encodes the hidden layer spin, the middle region II encodes the interaction with the visible layer spin, and the bottom region III encodes the magnetic field parameters experienced by the hidden layer spin. For the k-th wavelength incident on the phase-type spatial light modulator, the phase modulation of the three regions is as follows:

[0024]

[0025] Where L is the normalization parameter, and represents the maximum absolute value of the interaction and magnetic field parameters. This invention encodes each spin using a macroscopic pixel, which is composed of N... x ×N y It consists of modulator pixels, where m and n are pixel indices within each spin, 1 ≤ m ≤ N. x , 1≤n≤N y To achieve coded interactions in phase-type spatial light modulators, this invention proposes a canonical transformation: in region II, phase modulation also incorporates an additional contribution from checkerboard modulation, which is equivalent to rotating the spin of each visible layer by an angle α. ik =arccos(W ik / L) corresponds to the new spin v ik =v i W ik After the optical Fourier transform by the lens, the Hamiltonian measured by the detector is... Where Q is a constant.

[0026] Conversely, when the visible layer spin is obtained by sampling the hidden layer spin, the top region I encodes the visible layer spin, the middle region II encodes the interaction with the hidden layer spin, and the bottom region III encodes the magnetic field parameters experienced by the visible layer spin. The phase modulation of the three regions are as follows:

[0027] like Figure 4 The diagram illustrates the process of an optically confined Boltzmann machine generating an image. When the hidden layer spin is obtained by sampling the visible layer spin, the visible layer spin configuration v is randomly initialized. 0 The interactions and v obtained from training 0 The parameters are encoded into region 1 of the phase-type spatial light modulator according to formula (2). First, the spin value of all hidden layers is set to 1, and they are encoded into region 1 of the phase-type spatial light modulator according to formula (1). The magnetic field parameters obtained from training are encoded into region 1III of the phase-type spatial light modulator according to formula (3).

[0028] For each hidden layer spin h kThe probability of Gibbs sampling can be calculated from the difference in light field intensity detected twice. First, when the hidden layer spin is set to 1, the Hamiltonian measured by the detector is... To calculate the Gibbs sampling probability of the spin of the kth hidden layer, the spin of that hidden layer is independently flipped to -1, such as... Figure 4 As shown in the gray box, let h k =-1, the Hamiltonian measured by the detector is Therefore, the difference between the two detected Hamiltonians is At a specific temperature T, h is calculated using Gibbs sampling. k The probability of being 1 is Therefore, the method of this invention can reduce the computational complexity of Gibbs sampling from O(N) to O(1). By performing M optical Gibbs samplings between the visible and hidden layers until the obtained spin configuration is stable, the generated content is obtained from the final visible layer spin decoding.

[0029] For amplitude-type spatial light modulators, their optical path is similar to... Figure 1 Consistent. When the hidden layer spin is obtained by sampling from the visible layer spin, for the k-th wavelength on the amplitude-type spatial light modulator, the amplitude modulation of the three regions are respectively... Conversely, when the visible layer spin is obtained by sampling the hidden layer spin, the amplitude modulation of the three regions is as follows:

[0030] like Figure 5 The diagram shows the optical path using an electro-optic modulator. It includes a coherent broadband light source, single-mode fiber, a beam splitter, an electro-optic modulator, a beam combiner, a time-lensing system, and a photodetector. The beam splitter divides the coherent broadband light source into multiple wavelengths, each incoherent. For the k-th wavelength, the electro-optic modulator modulates the amplitude or phase of the light field over time; the time-modulated signal is consistent with the signal modulated along the y-axis of the spatial light modulator. Each modulated wavelength is combined into a single beam by the beam combiner, then passes through the time-lensing system. The intensity is detected by the photodetector, and finally, the computer provides feedback on the detected intensity and further modulates the light field.

[0031] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optically realizing a restricted Boltzmann machine, characterized in that, Based on a non-von Neumann architecture, a coherent broadband light source is used as the incident light. The light field is wavelength-splittered by a beam splitter and then incident on a spatial light modulator. The spatial light modulator is divided into three regions: the top region encodes the hidden layer spin, the middle region encodes the interaction and visible layer spin, and the bottom region encodes the magnetic field parameters of the hidden layer spin. The spin, interaction, and magnetic field parameters are encoded into the light signals at different locations by the spatial light modulator. An optical Fourier transform is performed through a lens system, and the intensity of the Fourier transform light is measured using a detector. Optical Gibbs sampling is achieved by calculating the difference between the two measured light intensities.

2. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, Able to achieve Hamiltonian as The restricted Boltzmann machine, among which visible layer spin, For hidden layer spin, visible layer spin With hidden layer spin The interaction between them visible layer spin Magnetic field parameters, Hidden layer spin The magnetic field parameters are given by K, where K is the total number of wavelengths.

3. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, Based on the spin, interaction, and magnetic field parameters of the visible and hidden layers, a spatial light modulator modulates optical signals at different locations, using either amplitude modulation or phase modulation. When all magnetic field parameters are zero, the spatial light modulator can modulate the optical signal solely based on spin and interaction.

4. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, Optical Fourier transforms are performed using a lens system, which includes ordinary lenses, micro / nano structured lenses, or time lenses.

5. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, After passing through the lens system, the light intensity after Fourier transform is measured using a detector, which includes a photodetector and an optical camera.

6. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, By calculating the difference in light intensity between two detections, the Gibbs sampling probability is calculated, reducing the computational complexity of Gibbs sampling to a minimum. .

7. The method for optically realizing a restricted Boltzmann machine according to claim 1, characterized in that, The optically constrained Boltzmann machine is used to generate content and perform classification recognition.

Images